Turkish PII masking model for Ollama and llama.cpp

turkish-pii-detection-GGUF is the GGUF build of turkish-pii-detection v02, a 270M model that masks Turkish personal data (names, national ID numbers, IBANs, phone numbers, emails, addresses; 53 labels) following a Turkish instruction. It runs locally and offline with Ollama or llama.cpp, without Python or Transformers.

Files: Q8_0 (300 MB), BF16 (551 MB)
Base model: halilneed/turkish-pii-detection v02
Benchmark: Q8_0 0.944, BF16 0.943 exact match on 1,000 rows
Runs in: Ollama, llama.cpp
Licence: Gemma Terms of Use

Files and scores

File	Size	Exact match, 1,000 rows	Same output as original
original model, Transformers BF16	536 MB	0.944	—
`turkish-pii-detection-BF16.gguf`	551 MB	0.943	99.6%
`turkish-pii-detection-Q8_0.gguf`	300 MB	0.944	98.9%

Measured on cagrigungor/turkish-pii-masking-benchmark with row-level exact match, the same metric as the original model. The GGUF files were run through Ollama 0.40.7 at temperature 0. Q8_0 has the same weak slices as the original (PII with suffixes 0.803, UPPERCASE text 0.847); the one difference is the 20-row multi-person slice, 0.650 against 0.750. If you mask texts that mention several people, use BF16.

Run it in Ollama

ollama pull hf.co/halilneed/turkish-pii-detection-GGUF:Q8_0

For deterministic output and the right stop token, create a named model with the Modelfile from the repository:

# Modelfile and .gguf in the same folder
ollama create turkish-pii -f Modelfile
ollama run turkish-pii "Metindeki tüm kişisel verileri uygun etiketlerle maskele.

Metin: müşteri Ayşe Yılmaz tc 12345678950 tel 0532 111 22 33"
# -> müşteri [AD] tc [TCKN] tel [TEL]

The prompt is always an instruction, a blank line, then Metin: and the text. The instruction sets the policy, for example “Sadece telefon numaralarını maskele” (mask only phone numbers).

Run it in llama.cpp

llama-cli -hf halilneed/turkish-pii-detection-GGUF:Q8_0 --temp 0 -n 512

Limitations

The original model’s limits apply: it was evaluated on synthetic data, so measure it on your own text before production; it can produce labels outside the schema, so validate the output against the label list. There is no Q4_K_M build. KVKK or GDPR compliance remains the user’s responsibility.

Questions

Which file should I use?

Q8_0 (300 MB) scores 0.944 row-level exact match on the 1,000-row benchmark, the same as the original model; BF16 (551 MB) scores 0.943. On the 20-row multi-person slice Q8_0 scores 0.650 against 0.750, so use BF16 for texts that mention several people.

Is there a Q4_K_M build?

No. 168M of the 268M parameters are in the embedding table and Q8_0 is already 300 MB. A 4-bit build is not published without measuring its loss.

What is the prompt format?

An instruction, a blank line, then Metin: followed by the text. The instruction sets the masking policy. Use the Modelfile in the repository for temperature 0 and the right stop token.

Does it guarantee KVKK or GDPR compliance?

No. It was evaluated on synthetic data; measure it on your own text and validate the labels against the schema. Compliance remains the user's responsibility.

Ollama ve llama.cpp için Türkçe kişisel veri maskeleme modeli

turkish-pii-detection v02’nin GGUF sürümü: Ollama ve llama.cpp ile yerelde, Python ve transformers olmadan çalışır. Q8_0 (300 MB) 1.000 satırlık benchmark’ta 0.944 tam eşleşme alır; çok kişili metinler için BF16 (551 MB) önerilir.